changes in poverty in rural ethiopia 1989-1995 ... · the last problem of using appropriate price...

Changes in Poverty in Rural Ethiopia1989-1995: Measurement, Robustness

Tests and Decomposition

Stefan Dercon and Pramila Krishnan

WPS/98-7

March 1998

Centre for the Study of African EconomiesInstitute of Economics and Statistics

University of OxfordSt Cross Building

Manor RoadOxford OX1 3UL

Stefan Dercon is at the Centre for the Study of African Economies and KatholiekeUniversiteit Leuven.Pramila Krishnan is at the Centre for the Study of African Economies.

Correspondence: e-mail: [email protected]

Abstract: Assessing changes in poverty levels over time is bedevilled by problems inquestionnaire design, the choice of the poverty line, the exact timing of the survey anduncertainty about the appropriate cost-of-living deflators. In this paper, we focus on testingthe robustness of measured changes in poverty to these common problems, using householdpanel data collected in rural Ethiopia in 1989, 1994 and 1995: in particular, we implement asimple graphical technique for assessing the impact of uncertainty in measured inflationrates. We find that poverty declined between 1989 and 1994, but remained virtuallyunchanged between 1994 and 1995. However, the last result disguises substantial seasonalfluctuations in 1994. We also find that households with substantial human and physicalcapital, and better access to roads and towns have both lower poverty levels and are morelikely to get better off over time. Human capital and access to roads and towns also reducethe fluctuations in poverty across the seasons.

1

1. Introduction

Identifying the pattern of change in welfare and poverty over time is of increasingimportance in the policy debate about reform in Africa. It is recognised that thereform programmes are only sustainable in the long run if they also result in povertyalleviation. However, the data available on changes in poverty in Africa issurprisingly limited compared to Asia1 . Despite the various household surveysrecently implemented (Deaton (1997)), problems ranging from access to data toincompatible surveys, have meant that few studies on the changes in welfare sincethe 1980s have been attempted2 . Cross-section data could be used to perform thistask, provided coverage and sampling are done with great care (Deaton (1997)).Panel data, although not without their own methodological problems are morereliable in establishing changes at least within the sample collected. In the context ofAfrica, with the exception of the rolling panels in some LSMS surveys, such as inCôte d’Ivoire (Grootaert et al. (1997)), the number of panel data sets that could beused for assessing the changes in welfare are limited. In this paper we use data froma survey conducted in 1989 in six villages in the Southern and Central part of thecountry. In 1994, these households were re-visited as part of a larger householdsurvey covering 15 villages throughout Ethiopia. Subsequently, the larger samplewas interviewed again in the second half of 1994 and in 1995. The result is a two-fold panel, the smaller one allowing the analysis of welfare changes between 1989and 1994, and a larger panel, covering 1994 and 1995.

The period analysed in this paper is ideal for such an exercise in the contextof reform in Ethiopia. The first survey, conducted in 1989, provides a picture of thesituation in Ethiopia towards the end of a long period of strict economic controls,bad weather and civil war3 . The year 1994 marks the beginning of a structuraladjustment programme, agreed by a new government that came to power after theend of the civil war in 19914 . Consequently, the smaller panel on about 350households can address change in the period after the end of the war and after thefirst wave of the reforms. The second panel (on about 1450 households between1994 and 1995) can be used to examine the initial consequences of the structural

1 For example, in India, there has been systematic and regular collection of the information neededfor an appropriate analysis of changes in poverty since the 1960s in the form of large cross-sectional surveys (Ravallion and Datt (1995)).2 Demery and Squire (1996) review six countries in which some attempt has been made to comparewelfare over time. Grootaert et al. (1995) and Grootaert and Kanbur (1993) analysed changes inCôte d’Ivoire between 1985 and 1988.3 The civil war, although started many decades earlier, had intensified in the 1980s with recurringoffensives by the government army and by the rebels in the Northern part of the country. Theeconomy had been brought to its knees after a period of an experiment with a strict control regime,ideologically inspired by close ties with the communist bloc of Eastern Europe. By 1989 resourceflows from the Soviet Union and other states had dried up after the collapse of the communistregimes in this region. Finally, the 1980s saw some of the worst famines ever in Ethiopia, inducedby drought and war.4 In 1990, the previous government embarked on partial reforms, with food-market liberalisation,the abolition of much of rural taxation and the forced supply of grain by peasants. Thegovernment was defeated in 1991 by a Tigrayan-led coalition which brought to an end theprotracted civil war. In 1992, the currency was devalued and in 1994, the new government agreeda programme of reforms and structural adjustment with the World Bank and the IMF.

2

adjustment programme5 . An initial analysis of the results of the smaller panelbetween 1989 and 1994 (Dercon and Krishnan (1994)) showed substantial declinesin poverty in some of the villages surveyed; in a few villages the decline was morelimited. The results have been used in Demery and Squire (1996) and in Jayarajah etal. (1996). The results were preliminary and in this paper we test the robustness ofthese results and extend the period of analysis. In general, we find that our previousfindings do stand. We observe an overall decline in poverty in the sample between1989 and 1995; this poverty decline is driven mainly by strong improvements insome villages, while in others little change is observed, and these results persistwhen controlling for seasonal effects. There is little change in measured povertybetween 1994 and 1995.

Measuring welfare changes is not without its problems6 . In this paper, weexplore whether the results obtained are robust to alternative solutions to some ofthe methodological problems. In line with most studies, we use consumption as ourbasis for measuring the standard of living7 . Furthermore, we use a cost-of-basic-needs poverty line to calculate poverty measures (Ravallion and Bidani (1994)). Themeasures used are from the Foster-Greer-Thorbecke family of additivelydecomposable measures (Foster et al. (1984)).

We focus on three problems: first, are our results sensitive to questionnairedesign; second, are they sensitive to the actual poverty line chosen (stochasticdominance) and third, are they sensitive to the sources of the price data used8 . Inparticular, we examine the consequences of potential errors in the measurement ofrural inflation in the survey sites. Of these problems, the first two have beendiscussed quite extensively (see Atkinson (1987), Deaton (1997), Ravallion (1994),Lanjouw and Lanjouw (1996)). In this paper, the discussion will be rather limited.The last problem of using appropriate price deflators, has been noted in some studies(Kanbur and Grootaert (1994), Ravallion and Bidani (1994)), although theconsequences for poverty measurement have not been systematically explored inintertemporal poverty analysis. In this paper, we present a simple dominance resultthat could fill this gap.

The paper is organised as follows. In the next section, we describe the dataused. In section 3, the construction of consumption and some of the problems in thedata related to compatible definitions of consumption are discussed. In section 4, thepoverty line used and the problems related to price information are analysed. Insection 5, we present the poverty findings and test the robustness using stochasticdominance. In section 6, the issue of the sensitivity to the measurement of pricechanges is discussed and a comprehensible and readily implementable methodpresented. Once the pattern of the changes in welfare is robustly established, thenext important issue is whether it is possible to explain these changes in the contextof the panel data. In section 7 of the paper, a simple poverty profile is described and

5 In 1997, a further round of surveys was completed, allowing a further comparison of changeduring the reform period, although the data are not yet available for this paper.6 For a discussion of some of the problems, see Lipton and Ravallion (1995)).7 Obviously, this is not without its critics, although there are good reasons to use it in practice(Anand and Harris (1994), Ravallion (1994), Hentschel and Lanjouw (1996)).8 Both the choice of the poverty line and the poverty measures have yielded by the largest literatureon the methodological problems in poverty measurement (Atkinson (1987), Ravallion and Bidani(1994), Ravallion (1994)).

3

a first interpretation of the factors explaining the changes between 1989 and 1995 isgiven. A more detailed analysis is the subject of future work. Section 8 concludes.

2. The data used

In 1989, the International Food Policy Research Institute conducted a survey inseven villages9 now located in the regions called the Amhara, Oromiya and theSouthern Ethiopian People’s Association. The study collected consumption, assetand income data on about 450 households. In 1994, the Centre for the Study ofAfrican Economies and the Economics Department of Addis Ababa Universitystarted a panel survey incorporating six of the seven villages earlier surveyed in 1989in its sample (the remaining village in a semi-pastoralist area in Southern Ethiopiacould not be revisited again because of violent conflict in the area). Nine additionalvillages were selected allowing for a total of 15 village studies, covering 1477households (the Ethiopian Rural Household Survey, ERHS). They were interviewedthrice: in the first part of 1994, again later in the same year and in the first part of1995.

In the 1989 survey, the households were randomly selected within eachcommunity, while the communities selected were mainly areas which had sufferedfrom famine in this period (for details see Webb et al. (1992), Dercon and Krishnan(1996)). Consumption information from the six villages surveyed in 1994 is availablefor 363 households. However, due to the extremely difficult survey conditions, dataon both food and non-food consumption were collected in only four villages (i.e. for213 households), while only food consumption data were collected in the other twovillages.

In 1994, the sample was expanded with nine additional communities, which wereselected to account for the diversity in the farming systems in the country, includingthe grain-plough areas of the Northern and Central highlands, the enset-growingareas and the sorghum-hoe areas10 . It is a self-weighting sample, with each personrepresenting approximately the same number of persons from the main farmingsystems. For 1994 and beyond we have complete data for most households (1411)for all three rounds. Within each village, random sampling was used, stratified byfemale headed and non-female headed households. In annex 1 we give details of thesampling method used. The resulting sample can be considered broadlyrepresentative of the households in the different farming systems in the country.Obviously, with only 15 communities, but relatively large samples within eachvillage, the interpretation of the results in terms of rural Ethiopia as a whole has tobe done with care. No other sources allowing a comparison over time exist,however, so that the current data set is probably the only one currently available tomake any statements about change in Ethiopia11 .

9 We use the term “village” in the paper for simplicity, although in fact the sampling unit is the“Peasant Association” , a formal administrative term describing one village or sometimes a smallnumber of villages, controlled by one administrative authority.10 The first round of the 1994 survey was conducted in collaboration with IFPRI, WashingtonD.C..11 The survey collected also extensive information on health and anthropometric outcomes of allpersons in the sample. In the same year, the Central Statistical Office collected a data set as part ofthe Welfare Monitoring System. Many of the average outcome variables, in terms of health and

4

An important issue for panel data is the attrition rate across rounds. Despitethe fact that the 1989 survey was not designed in order to start a panel householdsurvey, only 7 percent of households were lost in 1994. In most cases, this was dueto poor recording of names, rather than any systematic reason that could have biasedthe resulting sample. In 8 percent of cases, the head of the household had changed(due to death, illness or transfer of headship to a son or daughter because of age).These households were retained. Less than 2 percent of households were lostbetween the three rounds of the ERHS in 1994 and 1995.

Annex 1 gives more details about the survey sites. The survey was notconducted in exactly the same months in each round, so that comparison has to bedone with care. If seasonal consumption smoothing is less than perfect, for exampledue to variable food prices or imperfect credit and asset markets, then comparingdifferent survey years may reveal apparent welfare changes over time, which are infact due to seasonality. One simple way to avoid this problem is to compare resultson welfare using as closely related periods as possible. As can be seen in annex 1,this is not the exactly the same for all sites when comparing 1989 and 1994,although the first round of 1994 (referred to as 1994a) can be directly compared (interms of timing) with the third round (1995) for all villages.

3. Problems in questionnaire design and measurement issues

Several potential problems with comparing poverty over time exist and have beendiscussed in the literature. In this section we address the main problems related toquestionnaire design and the measurement of consumption. First, the problem ofchanges in questionnaires12 over different rounds of a survey needs to be addressed.Comparability is badly harmed with substantial changes in questionnaire design. Forthe 1994a, 1994b and 1995 round we do not have this problem since thequestionnaires were not changed. For 1989 data there is no fundamental problem:the 1994-questionnaire is modelled on the 1989 questionnaire, with all the mainitems prompted for in exactly the same way. The format of the consumptionquestionnaire is the same in all rounds: three questions on ‘did you purchase’, ‘didyou consume from own production/stock’, ‘did you consume from gift or wage inkind’, with lists of items for which the interviewee was prompted. However, thedifference between the 1989 and 1994 questionnaire was that the list of items usedin 1994 was slightly longer, since following piloting it was found that more itemswere commonly consumed than asked for in 1989. Questions on ‘did you consumeanything else’ were asked in all rounds, including 1989, so in principle the items notlisted or added as ‘other item’ were included in the 1989 survey. The fact that thelist was also shorter ex-post in 1989 than 1994 could simply be due to shortagesbefore the reforms and at the height of the economic crisis of the late 1980s.Nevertheless, as an additional check on the results, we recalculated the 1994 figuresusing only the items which were explicitly prompted for in 1989. We use the samepoverty line for both the limited and the expanded definition of consumption. By

nutrition were very similar to the results in the ERHS, suggesting that the resulting sample maywell be broadly representative of the general situation in rural Ethiopia. See Collier et al. (1997).12 Grosh and Jeancard (1994) and Lanjouw and Lanjouw (1997) discuss some of the consequencesif this were to happen. Appleton (1996) discusses the consequences for poverty comparisons inUganda.

5

limiting the items used in the calculations for consumption after 1994, we may wellbias the results against a reduction in the measured number of poor13 .

Another issue is the actual definition of consumption used. The actualconsumption definition used is the sum of values of all food items, includingpurchased meals and non-investment non-food items. The latter was interpreted in alimited way, so that contributions for durables and house expenses were excluded, aswell as health and education expenditures (see Hetschel and Lanjouw (1996)).Although there may be methodological reasons to so measure welfare in practice,excluding these items is also done to avoid further bias due to different prompting ofitems in 1989 and 1994. However, one would expect that since 1989, and the end ofthe war in 1991, households are spending more on durables or construction - assetswhich are typically risky investments in insecure times (Collier and Gunning (1996)).As a consequence, again, we may, if anything, bias the results against reductions inthe levels of poverty since 1989.

Another standard problem is related to the valuation of own production orgift consumption. We avoided the problem of using ‘within survey’ prices to valuethe very large consumption from own production or from gifts in kind (see Deaton(1989)). We collected data on prices in each village at the time of the consumptionsurvey itself14 . However, such a local price survey was not available in 1989. Ratherthan using unit values, we decided to focus on identifying an alternative sourcewhich could be used both in 1989 and 1994. A widespread price data-collectionexercise is undertaken every month by the Central Statistical Authority (CSA), butprices are reported at an aggregated level (e.g. only 4 prices are available for thevast SEPA region or Oromiya region). We also assess how using different priceseries would affect our findings15 .

Consumption data are available only at the household level so furthercorrections are needed. Households in developing countries often have fairlycomplicated structures. In annex 2 we briefly discuss the concept of the householdused, since several definitions were embedded in the questionnaire. Irrespective ofthe concept of the household, correcting for household size and composition is alsoan important issue. We calculated adult equivalent units using World HealthOrganisation (WHO) conversion codes. Since data on household size andcomposition was collected in each period, we adjusted the household size and theadult equivalent units in each period16 ,17 . In many respects, this remains a relatively 13 Lanjouw and Lanjouw (1996) suggest an alternative procedure for making poverty comparisonswhen consumption definitions differ.14 This proved more difficult than expected. Many items are not standard or available, even on thenearest urban market. These urban markets are often 5-10 km away and prices relevant for thehouseholds are not necessarily the same. Deaton (1997) reported that similar problems existed inmany of the LSMS-surveys. See also Grootaert and Kanbur (1994) for Côte d’Ivoire.15 A specific problem in Ethiopia was that no standard measures (kg, lt) are used by thepopulation. We identified about 100 different weights and measures, and to convert quantities, weconducted village-level conversion surveys. It was found that each village appears to have its owndefinition of commonly used measures, complicating our activities further. We also recalculated allthe consumption data from the raw 1989 data (questionnaires were checked again) to make surethat differences in the conversion factors used by the research team in 1989 were not responsiblefor any of our findings. Capéau and Dercon (1998) report on an alternative econometric approachto estimate prices and conversion factors. In that paper it is shown that the results obtained fromthe econometric approach and from the community level surveys are relatively similar, whilemethods using unit values provide a very different result.16 The equivalent scales used are in annex 3.

6

arbitrary correction, especially since consumption is not limited to just the intake ofcalories. Ravallion and Lanjouw (1995) provide a careful analysis of the robustnessof poverty measures to the weight attached to household size. This is beyond thescope of the current paper.

Table 1 provides means of total monthly consumption and food consumptionper adult equivalent for the 1994 round for the six villages for which data areavailable for 1989 as well. They are in birr per month (the official exchange rate atthe time was 5 birr per dollar). In the table, the comprehensive definition refers to afull list of items in 1989, while the limited definition includes only those items whichwere prompted for in the 1989 survey. The data in table 1 are for the 6 panel sitesusing the 363 observations. Note that we did not calculate the limited definitions forareas in which no equivalent data were collected in 1989.

Table 1 Consumption per adult equivalent in the panel sites in 1994:issues of definition

type ofconsumption

consumptiondefinition

Dinki DebreBerhan

AdeleKeke

Koro-degega

Garagodo Domaa

total comprehensive 67.5(73.9)

122.5(92.9)

161.7(137.9)

45.5(29.7)

30.9(26.3)

60.2(47.7)

limited 63.9(74.7)

119.2(92.5)

147.7(102.1)

n.a. 27.8(25.0)

n.a.

food comprehensive 58.9(70.2)

103.6(89.1)

119.8(114.6)

39.5**(26.0)

25.8(25.0)

52.6(46.8)

limited 55.0(70.0)

104.7(89.2)

105.1(80.6)

31.4(22.3)

22.4(23.4)

40.0(40.7)

Data from the 1994 round of the ERHS. All consumption figures are mean per adult equivalent inthe village, on average per month. Standard errors in brackets. Limited definition means that thelist of items explicitly prompted for in 1989 is used in 1994 as well. Comprehensive definition usesall data food and non-food consumption items recorded in the survey, excluding durables, healthand education.** = limited definition is significantly smaller than comprehensive definition at 1 percent or less.n.a. = not applicable, since no data in 1989

The differences on employing the alternative definitions do not appear very large.Only in one village is the difference in food consumption significant. Of course,these are mean values, not poverty measures. We investigate the consequences ofthe different definitions for poverty below.

17 For two villages in the 1989 survey no complete age profile of household members had beencollected. We only had numbers of male or female adults, and total number of female and malechildren under 15 years of age. We used the rest of the data to estimate the typical relationshipbetween adult equivalent units and the age-household structure as given by male and female adultsand children (i.e. aeu=f(male children, male adults, female children and female adults)). Theresults of the estimation were: aeu=1.04*male adults+0.80*female adults +0.76*malechild+0.69*female child. This regression was then used to obtain adult equivalent units in the twovillages with aggregate information only.

7

4. Constructing Poverty Lines to Analyse Changes in Poverty

The study of poverty in a country is ultimately an attempt to compare livingstandards across households or individuals. It therefore suffers from all the usualproblems associated with tastes, circumstances, price differences and behaviouralresponses. While economists may have little problem with using consumptionmeasures, one still needs to make careful corrections to allow monetary measures toreflect poverty differences. As usual, poverty will be defined relative to a povertyline. Although alternative methods to define the poverty line are possible (Anand andHarris (1994), Greer and Thorbecke (1986)), we use the cost-of-basic-needsapproach to estimate a poverty line (Ravallion and Bidani (1994)). A food povertyline is constructed by valuing a bundle of food items providing 2300 Kcal. A specificvalue for this basket is obtained per survey site. To this value, an estimated non-foodshare is added to obtain the total consumption poverty line per day per adult.

We identify two specific problems with this approach in the Ethiopian case.First, pricing a basic basket assumes the availability of all these commodities in thelocal market, which is difficult to believe especially for 1989. Indeed, weencountered problems with finding price data for some commodities in the localmarkets18 even in 1994. A second problem is that in rural areas we are dealing withvery different farming systems (enset versus cereal based systems, see annex 1).Their diets are very different, implying very different product availability in marketsaffecting our pricing. The main consequence of the latter problem appears to be verydifferent cost-of-living measures depending on which diet is used (specific per site orcommon for all sites). As discussed in Dercon and Krishnan (1996), the appropriateprocedure is not self-evident19 . In this paper we settled for a common diet foreveryone, to increase comparability across sites.

As will be seen below, the issue of prices becomes even more crucial whenattempting to do comparisons over time and space. We know from other work thatprice dispersion is high in Ethiopia, with markets taking considerable time toperform arbitrage (Dercon (1995)). Also, rural areas are not well served by ruralmarkets, probably due to very poor infrastructure, while even in small urban marketsthe availability is often poor. Even if markets always clear, price variability over timeis high, and is not explained by seasonal factors. Such variability is very difficult todeal with in analysing poverty. Temporary price increases will make the minimumfood basket very expensive, and the expected behavioural response is to reduceconsumption as long as prices are very high. When prices return to lower levelsconsumption may then be boosted. Depending on whether consumption wasmeasured when prices were high or low has important consequences for finding 18 These problems are common in this type of survey. See Deaton (1997) and Capéau and Dercon(1998) for a discussion and some alternative solutions.19 The problem is linked to the issue of compensation for needs versus tastes (Ravallion and Bidani(1994). If it is clearly a matter of choice that in some areas, such as urban areas, householdsconsume more expensive commodities, then compensation for these expensive tastes is unlikely tobe appropriate in rural-urban poverty comparisons. However, the differences in diets in Ethiopiaare closely linked to farming systems that have developed over very long periods. This may suggestthat a specific poverty line for each system or village in the survey may not be inappropriate. InDercon and Krishnan (1994), it is shown, however, that this reverses the order of villages in termsof poverty. In this paper, we are dealing with changes over time, and it was found that the patternof change is hardly affected by this discussion, so that we settled for the simple common povertyline (in terms of the quantities included in the diet, not in terms of its value) for all sites.

8

whether households were poor or non-poor. In fact, since allowing consumption tofluctuate may be part of the same consumption plan, the interpretation of thepoverty figures is difficult: when prices are (temporarily) high, poverty is likely to beoverestimated, while when prices are low, poverty is likely to be underestimated20 .Seasonality presents a similar problem, but here, information about the likelypatterns of prices is available since the seasons are always with us.

W decided to use the same basket of commodities for each period and site toincrease transparency and comparability in the analysis, using 1994 as a base year todetermine the basket of commodities included21 . As in Ravallion and Bidani (1994),we constructed a typical diet for the poorest half of the sample in nominalconsumption using the 1994 data and calculated its calorie contribution22 . We thenscaled this measure to reach 2300 Kcal per day. The diet is given in annex 3, tableA.5.

We used the approach described in Ravallion and Bidani (1994), to estimatethe required non-food share by estimating an Engel curve and then determined thefood share of the representative household whose total consumption is exactly equalto the food poverty line. Details are given in annex 423 . The value of the non-foodshare at the poverty line can then be interpreted as representing the absoluteminimum basic needs in terms of non-food items, for which households should becompensated, on top of the minimum food requirement. The resulting food share atthe poverty line is 83 percent on average. Note that this share is very high, so thatthe non-food share to be added to the food poverty line is actually quite low. Theconsequence is that this implies that the total consumption poverty lines calculated inthis way are relatively low. Indeed, they are close to 10 dollars per month per adult,much lower than one would find in many other African countries.

A few remarks on this ‘low’ poverty line are in order. Although the approachaims to establish an ‘absolute’ poverty line by measuring the actual cost of basicneeds, its application does not necessarily result in a poverty line that could bedirectly used for comparisons across countries. We use data from the survey itself todecide the relevant minimum food bundle to establish the poverty line. In doing so,we limit it to calorie-intake. Of course, calorie-intake is only a limited part of ahealthy diet; if a large part of the country is then to perforce forego other moreexpensive nutrients to obtain a calorie-intensive diet, then the resulting diet to reach2300 Kcal is biased against the inclusion of other nutrients. If other nutrients wereincluded in the construction of the diet, then we would probably have reached amuch more expensive food diet. For example, the only protein intake included isfrom pulses and milk; no meat or fish is included, since the poorer half of the sample

20 If the problem is mainly intertemporal variability, a possible solution is to make the minimumbasket of commodities dependent on the time period - effectively adjusting over time the quantitiesneeded to obtain the minimum level of consumption. If the variability is mainly spatial then onemay argue in favour in taking location-specific diets. However, this raises again the problem ofcomparability.21 The poverty line then effectively becomes a cost-of-living index with budget weights taken fromthe poorer half of the sampled households.22 Fortunately, all these commodities were prompted for in the 1989 survey as well.23 One could argue that non-food shares could be calculated for each site separately. However, ifimplemented in this way, this would only have been appropriate if they reflected genuinedifferences in needs or relative food/non-food prices across areas. Since we could see little groundfor such an approach in our survey villages, and given the relatively small samples within eachvillage, one non-food share was used for all areas.

9

simply do not consume it. Since food shares decline with total expenditure, non-food shares near these new food poverty lines with more nutrients would also behigher, resulting in an even higher total poverty line. An important consequence isthat the poverty measures calculated in this way can hardly be used for cross-country comparisons; for such comparisons, one-dollar-a-day or similar approachesmay be more appropriate.

The poverty line used for each period uses the same basket throughout, butvalued at the prices for the survey period. The poverty line can therefore also bethought of as a price deflator allowing comparisons across villages and over time. Apotential problem is that in 1989 and 1994 we are forced to use different pricesthrough lack of a specific price survey in the survey area during 1989, while duringthe three rounds since 1994, a site-specific price survey was collected. The regionalprice data from the CSA (Central Statistical Authority) are the alternative available.Since the CSA collected similar data in 1989, 1994 and in 1995, in the same periodas the rural survey, we use their data to value the minimum food basket for thesethree periods24 . This will give us a means of checking whether the price data sourcesmatter for the poverty comparisons over time.

In annex 5 we give the poverty lines for each site for 1989 to1995 for the sixpanel sites and for all sites in 1994 and 1995. In table 2 and 3, we give the averageof the poverty lines used both for the longer and the shorter panel. We also expressthem as an index to compare it with other data sources on price changes.

Table 2 Poverty lines and implied inflation rates : panel sites onlyPovertylineERHS

PovertylineCSA

PriceindexERHS

PriceindexCSA

PriceindexCPI

PriceindexFood CPI

1989 22.3 (100)* 100 100 1001994a 49.2 44.2 221 198 175 1851994b 48.3 216 184 1971995 50.8 48.0 228 215 180 193*using CSA 1989 =100 as baseSources: ERHS = price survey of the Ethiopian Rural Household Survey; CSA = regional price databased on Central Statistical Authority price data collection; CPI = official Consumer Price Indexbased on urban price data; Food CPI = food Consumer Price IndexPoverty lines for ERHS and CSA data are population weighted averages within the sample.

In table 2, the first two columns provide comparisons of the poverty line using theERHS price survey, compared to the regional data from the CSA. The 1994apoverty line is 11 percent higher when using the ERHS data. Since, for the 1989poverty line, we use the CSA data, we may overestimate the increase in the cost ofliving between 1989 and 1994 if we were to use the ERHS data for the latter period.Note that this difference is perfectly plausible, given the different markets in whichprices were collected. The CSA data include many rural market towns, while oursample specifically uses the local market, closest to the village, which in some casesis quite remote.

The differences between these two data sources become relatively small,however, when comparing the results with the situation using the CPI (official 24 We do not have equivalent data from the CSA coinciding with the second round of the ERHS,since all activities of the CSA were suspended at the time due to the 1994 Census.

10

Consumer Price Index) data. Irrespective of whether we use the overall or the foodCPI, both the ERHS and to a lesser extent the CSA price data suggest much largerprice increases between 1989 and 1994 than the official CPI. This points to thedangers if no careful choices are made with respect to price data: if we were to makepoverty comparisons simply using the CPI as the appropriate adjustment of the cost-of-living over time, then we are likely to underestimate the cost of basic needs, i.e.underestimate the level of poverty in our sample in 1994, in comparison to 1989.Part of the reason is likely to be the fact that the CPI is based on urban data only.

We looked for other means of checking the results. A possibility is to esti–mate poverty lines without price information. Greer and Thorbecke (1986) use suchan approach. We estimate a variant of their model. By regressing the logarithm ofcalorie consumption per adult equivalent on the logarithm of food consumption peradult equivalent, one is able to find the level food consumption that implies in thedata the consumption of 2300 Kcal per day per adult equivalent. Estimation withfood rather than total consumption was done because of the limitations on the dataavailable for 1989 (see section 2). We then calculated the value of food consump–tion at which the poverty line of 2300 Kcal per day was consumed. We findremarkably close estimates of the food poverty line to those calculated by the otherapproach: 20.7 birr in 1989 and 41.3 birr in 1994 (for comparison: the average foodpoverty line underlying table 2 is 18.5 birr for 1989, while for 1994, 36.7 birr usingthe CSA data and 40.7 birr using the ERHS price survey). These estimated foodpoverty lines suggest a 99 percent increase in nominal terms since 1989 - virtuallythe same as in the CSA rural prices, but higher than the CPI price increases. Thisappears to confirm the problems related to using the CPI for rural price changes.The level of the estimated food poverty line in 1994 is however closer to the foodpoverty line using the ERHS, suggesting that the ERHS price survey is the moreappropriate absolute measure of the cost of living to reach consumption levels closeto the poverty line. However, since we are especially interested in measuring thechange in poverty as accurately as possible, it appears more appropriate to use theCSA price data for the poverty line in both 1989 and 199425 . In sections 5 and 6, welook at the consequences of using different price sources.

Table 3 Poverty lines and implied inflation rates : all sitesPovertylineERHS

PovertylineCSA

PriceindexERHS

PriceindexCSA

PriceindexCPI

Price indexFood CPI

1994a 44.5 44.4 100 100 100 1001994b 47.0 106 105 1061995 50.0 47.6 113 107 103 104

sources: see table 2

Table 3 highlights another potential problem. Using the ERHS price survey, weobserve much larger price increases between 1994 and 1995 than those implied bythe CPI during exactly the same period: the ERHS data suggests a 13 percentincrease, while the CPI suggest only a 3 percent rise. The CSA regional prices

25 An alternative would be to impute a poverty line for 1989 from the 1994 food poverty line usingthe ERHS data and using the inflation rate in each site implied by the CSA data. This impliesadditional imputation, possibly causing further measurement error.

11

increased less than the ERHS, but still more than the CPI. Again, this illustrates theproblems with using the CPI within the rural sample as a means of adjusting thepoverty line over time.

5. Poverty levels and changes

Having constructed poverty lines and consumption measures of welfare, we can nowanalyse levels and changes in poverty. First, we focus on the panel households forthe trends between 1989 and 1994. Recall that for four villages, we have data ontotal consumption for both 1989 and 1994. For the six villages (and 361 households)surveyed, we have data only on food consumption in 1989 for comparison with1994. We construct food poverty levels using the full sample and total poverty levelsfor the 211 households with only food consumption data in 1989. Next, we look atthe pattern since 1994 as well. From 1994 onwards, we have a full panel withrelatively little attrition (see Annex 1). By 1995, the sample consists of 1411 withfull information in all three rounds for our purposes.

The poverty measures reported are from the FGT-family of poverty indexes(Foster et al. (1986)). Let yi denote consumption per adult equivalent which isordered for all households from low to high, and z the poverty line and if there are qhouseholds with consumption per adult below the poverty line z, then the Pα familyof poverty indexes can be defined as:

Pα = (1/n). ∑I=1q ((z-yi)/yi)α

(1)

for different values of α: if α = 0, this is the head count index, α=1 is the povertygap and α=2 is the severity of poverty index.

Since poverty measures are calculated using sample data, it is important to treatthem as statistics26 . We report poverty levels for households, not at the level of theindividual. Often poverty is reported by individuals by using the household sizes toconvert the household level observations in apparent individual level data. We donot follow this practice, because it artificially makes it appear that the sample size is 26 Kakwani (1990) provided standard errors and showed the conditions under which differences

between poverty measures are asymptotically normally distributed. He shows that standard error

(SE) of the difference between the estimates of two independent poverty measures P1* and P2

* is

equal to:

SE(P1* - P2

*) = (σ*1/n1 + σ*2/n2)1/2 (2)

in which σ*1 and σ*2 are the sample estimators of the variances of the asymptotic distributions of

n1(1/2). P1

* and n2(1/2). P2

*.The test-statistic for testing equality of the two measures:

η = (P1* - P2*)/ SE(P1* - P2*) (3)

follows an asymptotic normal distribution with zero mean and unit variance. He shows that thevariance of the asymptotic distribution of each estimated poverty measure of the Pα - family equals:

var (n1/2 .Pα*) = (P2α

* - Pα*2) (4)

12

much larger than actually is the case. This is important when calculating standarderrors of the poverty measures, as in Eqn (4) (see footnote): the larger the samplesize, the lower the error and the levels and differences will more often besignificantly different from zero. By using the data as if the number of times eachhousehold’s consumption level appears in the data is equal to the number ofhousehold members, the formula for the variance in (4) is not correct, since it doesnot take into account the extensive clustering implied by using the household as thesample unit, and not the individual. In principle, we could correct for this problem bycalculating the corrections for clustering (see Deaton (1997), Howes and Lanjouw(1996) for details), but this is beyond the scope of this paper. To investigate the robustness of the results relating to the change between 1989 and1994, we use two different definitions of consumption for 1994: the comprehensiveand the limited definition discussed in section 3. We also use two different povertylines: one using the CSA prices and one using the ERHS price survey data collectedin the sample villages. Table 4 reports food poverty level for the full panel (sixvillages) between 1989 and 1994.

Table 4 Food poverty levels 1989-1994; 6 panel villages (n=361)1989 1994a - ERHS

prices &comprehensivedefinition

1994a - ERHS prices &limiteddefinition

1994a - CSA prices &comprehensivedefinition

1994a - CSA prices &limited definition

P0 61.2 49.0 (-3.32) 58.2 (-0.83) 44.6 (-4.54) 52.1 (-2.49)P1 29.3 20.6 (-4.09) 26.5 (-1.28) 18.2 (-5.39) 23.3 (-2.77)P2 17.5 11.2 (-4.21) 15.4 (-1.29) 9.8 (-5.16) 13.6 (-2.40)

ERHS= poverty measure using poverty line valued at ERHS price survey;CSA = poverty measure using poverty line valued at CSA regional price survey;comprehensive definition = food consumption per adult using all items recorded in 1994;limited definition = food consumption per adult only using items prompted for in 1989.In brackets, the t-test statistic for testing differences in levels of poverty with 1989.The standard errors of each measure are not reported, but each was significantly different fromzero.

Looking at the results, it is obvious that poverty levels in 1989 in these villages werevery high, with a head count index of 61 percent. Using all food consumption itemsrecorded in the questionnaire in 1994 and using the local price survey collected atthe time of the survey, we find a large and significant decline in poverty. The headcount declined by a fifth and the intensity of poverty index by a third. Thesubsequent columns investigate whether the particular method of calculatingconsumption and the use of the ERHS price survey in 1994 and CSA prices in 1989,affects the results. Since the ERHS prices appear to suggest larger price increasessince 1989 than the CSA data, it is obvious that in that case the poverty decline issmaller. Similarly, by excluding some values for consumption items from the foodconsumption estimate, poverty is increased. Note however that poverty still declines:only if both the relatively high ERHS prices and the lower consumption estimatesare used is the consumption decline insignificant. If we use the same definition forconsumption and the same (CSA) source of prices for both 1989 and 1994, then thepoverty measures decline by 15 to 22 percent, depending on the measure. However,this decline hides the differences in experience across the different villages in thesample. Table 5 gives details for food poverty levels in 1989 and in 1994, using on

13

the one hand the full data and prices from the 1994 survey, and on the other handthe same data and definitions as in 1989.

Table 5 Food poverty levels 1989-1994 - panel villages

Dinki Debre Berhan Adele Keke Korodegaga Garagodo DomaaP0 89 41.5 33.9 41.9 74.7 80.0 84.9P0 94(1)

47.2 (0.59) 19.4 (1.85) 14.0 (3.04) 57.9 (2.50) 85.5 (0.76) 60.4 (2.95)

P0 94(2)

45.3 (0.39) 16.1 (2.33) 4.7 (3.04) 68.4 (0.97) 90.9 (1.64) 62.3 (2.74)

P1 89 14.4 11.8 10.4 39.7 45.9 44.2P1 94 (1) 16.2 (0.43) 5.0 (2.35) 3.8 (1.95) 21.8 (4.56) 47.2 (0.25) 28.5 (2.91)P1 94 (2) 15.3 (0.21) 2.4 (3.52) 4.3 (1.75) 25.8 (3.44) 58.3 (2.36) 30.5 (2.46)P2 89 6.6 5.4 4.7 24.8 30.2 26.5P2 94 (1) 7.2 (0.24) 1.6 (2.47) 1.4 (1.53) 10.4 (5.02) 29.4 (0.19) 16.7 (2.33)P2 94 (2) 6.6 (0.03) 0.5 (3.33) 1.9 (1.30) 13.1 (3.90) 40.0 (2.12) 19.1 (1.63)obs. 53 43 62 95 55 53

note: 94 (1) = poverty measure using poverty line valued at ERHS price survey and consumption per adult usingcomprehensive definition of consumption;94 (2) = poverty measure using poverty line valued at CSA regional price survey and consumption per adult usingdefinition of consumption, limited to items explicitly included in 1989 survey.In brackets, t-test of difference of estimate with the estimates in 1989.

In two villages we observe increases in food poverty, while in the others we observesubstantive decreases in poverty. The increases in Dinki are not significant, but thosein Garagodo are, for the poverty gap and the intensity of poverty, provided we usethe limited consumption definition and the CSA prices. In the other villages, thedecreases are generally significant for all measures and for the different methods27 .

Stochastic dominance tests provide further robustness tests of theconclusions about the changes in poverty. Atkinson (1987) discusses the relevantconditions to apply dominance tests for poverty measures28 . Figure 1 gives theappropriate cumulative distribution for 1989 and for two definitions of consumptionand price sources for 1994: consumption using the comprehensive definition withthe ERHS price data, and consumption and poverty lines using the same definitionand source of price data for both periods, i.e. the limited definition with CSA prices.We use the food poverty data for panel households in the six villages. The figuredemonstrates that everywhere, for a very wide range of poverty lines, the alternative

27 For two villages, Domaa and Korodegaga, we do not have non-food consumption data. In bothvillages we observe important decreases in food poverty, so it should not be a surprise that if weestimate for the four remaining villages total poverty estimates, we find insignificant changes. Intwo villages poverty increases, in the two remaining villages, poverty decreases. Overall, in thesample of 211 households for which we have total poverty estimates, poverty marginally increasesfor all poverty measures, poverty marginally increases for all poverty measures. The head countindex using the same definitions and price data sources in both years goes from 39.8 to 41.7percent; the poverty gap from 17.1 to 19.1 and the intensity of poverty index from 10.1 tot 11.5percent.28 For the FGT-poverty measures used in this paper and any other monotonic transformation of anadditive measure, First-Order Dominance can be defined for a particular range of poverty linesfrom 0 up to, say, z+. The condition states that poverty between two periods has unambiguouslyfallen if the poverty incidence curve for the latter period lies nowhere above that for the formerperiod within the range defined by 0- and z+. In our context, the poverty incidence curve is thecumulative distribution of households over different levels of consumption per adult. Since thedifferent distributions over time have to be put in the same graph, consumption per adult needs tobe expressed in comparable units, which is possible by defining them as multiples of the povertyline in each period (so that at the original poverty line as defined in table 2, real consumption isequal to one).

14

definitions for 1994 have little influence on the curves, and everywhere, the 1994poverty incidence curve is well below the 1989 curve. First-Order Dominancetherefore applies for all reasonable food poverty lines. Note that this means that alsofor higher order Pα measures, food poverty will be unambiguously lower in 1994(Atkinson (1994)).

Figure 1Stochastic Dominance Food Poverty 1989-1994

0%

10%

20%

30%

40%

50%

60%

70%

80%

90%

0.1

0.3

0.5

0.7

0.9

1.1

1.3

1.5

1.7

1.9

food consumption as multiples of poverty line

cum

ula

tive

per

cen

tag

e o

f h

ou

seh

old

s

1989 1994, ERHS prices 1994, CSA prices

Thus far in this section, we have only concentrated on the households in the samplefor which data exist in 1989. The ERHS household survey for 1994-1995 has moreextensive coverage and data were collected thrice over the year. The data in 1995were collected in more or less the same month as in the first round of 1994 (1994a).Therefore, they provide a test whether a year later, any change has occurred in thesample. The second round of 1994 (1994b) provides an interesting test on whetherthe exact timing of data collection matters for these welfare comparisons over time.In other words, seasonal effects can be captured. Table 6 presents the results for thePα measures. In brackets, we give the t-values of the test in the difference in theestimated poverty measure with the equivalent measure in 1994a. In annex 6, wegive the same table for food poverty levels (table A.10). The results are very similarin either case.

15

Table 6 Poverty levels 1994 - 1995 - ERHS panel householdsNorthernCereal

Central Cereal SouthernCereal

Southern Non-cereal

All Areas

P0 1994a 32.5 23.1 32.2 46.9 34.1P0 1994b 23.1 (-2.53) 14.3 (-3.26) 26.7 (-1.46) 41.8 (-1.52) 26.9 (-4.14)P0 1995 28.7 (-1.00) 23.3 (0.08) 28.8 (-0.90) 55.9 (2.62) 35.4 (0.71)P1 1994a 11.6 6.8 13.6 19.6 13.0P1 1994b 6.1 (-3.63) 4.0 (-2.79) 7.6 (-3.60) 13.9 (-3.39) 8.2 (-6.40)P1 1995 11.2 (-0.20) 6.7 (-0.13) 8.9 (-2.73) 24.0 (2.36) 13.3 (0.28)P2 1994a 5.9 2.9 7.3 11.1 6.9P2 1994b 2.4 (-3.76) 1.9 (-1.69) 3.2 (-3.95) 6.7 (-3.84) 3.7 (-6.48)P2 1995 6.0 (0.06) 2.8 (-0.09) 4.0 (-3.24) 13.1 (1.56) 6.8 (-0.15)

n 286 407 292 426 1411Notes: Northern Cereal are villages located in the Northern Highlands grain-plough complex;Central Cereal are villages located in the Central Highlands grain-plough complex; SouthernCereal are the villages in the grain-plough areas of Arsi/Bale or with sorghum plough/hoe; theSouthern Non-cereal are the enset villages with or without coffee/cereals. For details see table A.1and A.2. In brackets, the t-values testing the difference in the estimate of the poverty measure inthe particular period with the estimate in 1994a.

In terms of the full sample, there is a large and significant decrease in povertybetween the first and second round of the 1994 survey: poverty decreased by a fifthin terms of the head count and with even larger declines in the higher ordermeasures. The results for 1995 illustrate, however, that this is most likely to be astrong seasonal effect. Although there are differences between many areas in theexact timing of harvests, in the majority of the areas, the second round is thebeginning of the harvest in most cereal areas, when food is relatively plentiful. Thefirst (1994a) and the third round (1995) were conducted several months past themain harvest in most of these sites. Overall, we cannot detect a significant changebetween 1994a and 1995: aggregate poverty appears not to have been affected bythe reforms initiated 1994, at least in the short run.

As is to be expected, this obscures some differences between areas. In allareas, the decline in poverty between 1994a and 1994b is observed, and virtually inall cases it is significant. Only in the Southern cereal areas do we observe asignificant decline in poverty between 1994a and 1995, while in the Southern non-cereal villages we observe a significant increase. A tentative explanation for thelatter effect is that this is largely due to an increase in enset pests destroying somecrops in one village and a large decline in the possibility of seasonal migration due toethnic conflict in another village, which affected slack season earnings substantiallyin the area.

These results may well be specific to the actual poverty lines chosen. Tocheck the sensitivity of the results to different poverty lines, we show the results oftesting for stochastic dominance by plotting as before the cumulative distribution ofhouseholds under the poverty line for different multiples of the poverty line in eachperiod (figure 2). It can seen that these poverty incidence curves for 1994a and 1995(the first and the third round, collected at roughly the same months) are barelydistinguishable, confirming very similar poverty levels in both periods. Note that thecurves appear to cross a few times, suggesting the absence of first order stochasticdominance between 1994a and 1995. In principle, we could look into second orhigher order dominance by plotting the curves resulting from the integration of theseincidence curves (Atkinson, 1987; Ravallion, 1994). However, given that thedifference in poverty is sufficiently small as never to be significant, we did not

16

conduct these tests. From figure 2, we observe that the difference between the twoperiods in 1994a and 1994b is very large irrespective of the poverty line, suggestingconsistently significant seasonal differences in poverty between these two periods.Since it was found to be valid for all poverty lines larger than zero, first orderdominance applies, so that poverty for all measures considered is unambiguouslylower in 1994b compared to 1994. This confirms that seasonality may well affectany attempt to measure changes in poverty over time very considerably. Moststudies do not consider these problems when comparing survey data over time.

Figure 2 Stochastic Dominance Poverty Incidence Curve 1994-1995

0%

10%

20%

30%

40%

50%

60%

70%

0.40

0.50

0.60

0.70

0.80

0.90

1.00

1.10

1.20

1.30

1.40

1.50

1.60

1.70

1.80

1.90

consumption as multiples of poverty line

cum

ula

tive

per

cen

tag

e o

f h

ou

seh

old

s

1994a (first round) 1994b (second round) 1995 (third round)

Table 7 presents the levels of food poverty, following the initial sample in 1989 overthe three subsequent rounds. The results are broadly similar, even though the actuallevels of poverty from 1994 onwards are higher than in the full sample. This reflectsthe fact the 1989 survey focused on villages which had suffered from the drought inthe 1980s, making them poorer than the average village. Note that the seasonaleffect in these villages is even larger than in the full sample. The decline in povertybetween 1994a and 1995 is not significant but the observed decline between 1989and 1995 is substantial: about a quarter lower for the head count, a third less interms of the poverty gap measure and 40 percent lower in terms of the severity ofpoverty index.

Table 7 Changes in food poverty 1989-1995 - panel households (n=351)1989 1994a 1994b 1995

P0 61.3 49.6 (-3.13)a 33.3 (-4.51)b 45.3 (-1.13)b

P1 29.2 21.1 (-3.82)a 11.8 (-5.19)b 18.6 (-1.27)b

P2 17.4 11.4 (-3.93)a 5.9 (-4.52)b 10.3 (-0.78)b

a=t-test of difference with poverty in 1989b=t-test of difference with poverty in 1994aNote that the results deviate marginally from table 4 since some additional households were lost in1994b and 1995, compared to 1989.

17

Since seasonality of poverty looms large in these data, it is worthwhile to construct amore careful comparison in poverty between 1989 and 1995. Given that the exactdates of data collection differ between 1989 and the first round of 1994 or the 1995data in some areas, we constructed a new comparison, taking the closest month ofdata collection in the 1994-1995 rounds to make the relevant comparison with the1989 poverty levels. Table 8 gives the results. As might have been expected fromtable 7, in table 8 we still observe a large decline in poverty since 1989.Nevertheless, the observed decline is not a pure seasonal effect: poverty declinedsubstantially between 1989 and 1995 in this sample29 .

Table 8 Changes in food poverty 1989-1995, controlling for seasonality (n=351)1989 1994-1995

P0 61.3 45.9 (-4.14)P1 29.2 18.4 (-5.17)P2 17.4 9.9 (-4.88)

note: (1) Debre Berhan and Dinki = 1994a; Garagodo and Domaa =1995; 1994b=Korodegagaand Adele Keke.(2) T-test for difference in poverty measure in brackets.

6. Sensitivity of Poverty Measures to Price Changes

The results on changes in poverty in the previous section were derived using theERHS price survey, with some testing of the sensitivity if the CSA regional pricedata were used instead. There is little difference in using either source of price data.However, both data sources predict relatively larger price increases than some otherprice data sources. For example, the CPI increased by only 3 percent between 1994aand 1995, while on average the ERHS data suggest an increase by 13 percent (table3). Estimated price increases between 1989 and 1994a also differ by the datasources.

It is likely that these problems arise in many other contexts. In the LSMS-data on the Côte d’Ivoire, for example, these concerns about price data have givenrise to several different price index estimates. Grootaert and Kanbur (1994) havedocumented the sensitivity of the poverty measures to these results, by calculating alarge range of different measures. In practice, this is very time consuming and attimes no alternative data are available to cross-check a constructed price index withalternative price sources. Ideally, one would like to have some dominance results interms of ranges of inflation rates over which one can confidently predict that thepoverty orderings over time (or across space) remain the same.

Standard stochastic dominance tests do not allow for this problem.Effectively, the dominance results as in Atkinson (1987), are for povertycomparisons with a common poverty line for “real” consumption, i.e. consumptionvalues comparable over space or over time. In this paper we use the equivalent 29 Of course, these results hide differences between the experiences in different villages. In twovillages (Korodegaga and Domaa) there are large and significant declines in poverty using themeasure correcting for the seasonality in poverty. In two other villages (Debre Berhan and AdeleKeke), the decline is generally not significant, while in the two remaining villages there areincreases, although in one village (Dinki) they are not significant at all, and in the other village(Garagodo) only significant at 5 percent.

18

formulation of poverty defined over nominal consumption, with the povertymeasures corrected for price changes. It is readily seen that for all poverty resultsbased on the normalised poverty gap, (z-y)/z, such as the measures defined by (1),this is equivalent to using a constant poverty line z and a price deflator pt. Formally,let zt = pt.z. The normalised poverty gap can then alternatively be written as (zt-y)/ztand (z-y/pt)/z.

Stochastic dominance tests in poverty analysis checks whether the povertyordering remains the same over different multiples of the poverty line. Formally, theyinvestigate poverty ordering for different poverty lines defined as k.zt, in which k isscalar which is varied and zt is a specific poverty line in period t. Despite the factthat k is allowed to vary, the ratio zt+i/zt is kept constant across the poverty curves.Consequently, none of the dominance results obtained illuminate the problem of thesensitivity to uncertainty about the relative poverty lines. It is possible to derivesome dominance results in this context (see e.g. Jenkins and Lambert (1997)). Wepropose however a very simple graphical technique to illustrate the robustness of theresults to the analyst (for a formal discussion, see Dercon (1998)).

The basis of the technique is to define the estimated price change (inflationrate) at which poverty level are the same in the two periods30 . In principle this couldbe calculated for any poverty measure. Here we illustrate it for the head count. Infigure 3 we give the inflation rates between 1989 and 1994 for the 6 panel villageswhich would have made poverty in terms of the head count the same in each period,for different values of the head count. To calculate this line, we first ordered nominalconsumption per adult and constructed the cumulative distribution of households foreach level of nominal consumption. By looking at this distribution, we could, foreach cumulative percentage of households, determine the corresponding nominalconsumption levels in each period. Since the head count in simply the percentage ofhouseholds with less than a particular level of consumption, the ratio of the nominalconsumption levels obtained in each period would give the price deflator needed toput exactly the same number of households under the poverty line in either period.This deflator, expressed as a percentage change between 1995 (1994a) relative to1989 is given in figure 3. If actual inflation is higher than this figure, then povertywould have increased between 1989 and 1994-95; if it is lower, then a decline wouldhave been observed.

30 The approach can be easily adapted to problems of relative prices across localities.

19

Figure 3 Inflation rates for equal poverty 1989-1995

0%

50%

100%

150%

200%

250%

300%

15%

18%

22%

26%

29%

33%

37%

41%

44%

48%

52%

55%

59%

63%

66%

70%

74%

77%

81%

85%

poverty incidence

infl

atio

n r

ates

req

uir

ed

betw een 1989 and 1994a betw een 1989 and 1995

Recall that table 2 showed that the ERHS data suggest an average inflation rate of121 between 1989 and 1994a and 128 percent between 1989 and 1995. Inflationwas estimated to be much lower by all other sources. Equal poverty would haverequired a poverty rate well above 150 percent - and at higher levels of the povertyline, even larger than 200 percent. Hence, for poverty to have remained equalbetween 1989 and 1995, even higher inflation levels would have had to apply.Clearly, the observed changes in the six villages are robust even to substantialunderestimation of inflation in the ERHS survey. Problems with the inflation ratesare unlikely to matter here.

The changes between 1994 and 1995 can be analysed in a similar way.Figure 4 gives the results, comparing 1994b with 1994a and also 1995 with 1994a.This allows an analysis of the sensitivity of the results to the price factors in theseasonal and the one-year changes. Again, if inflation figures are higher than thoseimplied by the curve, then measured poverty is bound to have increased; if they arelower, a poverty decrease is likely to be observed.

20

Figure 4 Inflation rates for equal poverty 1994-1995

-10%

-5%

0%

5%

10%

15%

20%

25%

30%

35%

15%

19%

23%

26%

30%

34%

38%

42%

46%

49%

53%

57%

61%

65%

69%

72%

76%

80%

84%

poverty incidence

infl

atio

n r

ate

betw een 1994a (round 1) and 1994b (round 2) betw een 1994a (round 1) and 1995 (round 3)

The top curve, showing the inflation rates between 1994a and 1994b needed to keeppoverty equal illustrates the robustness of the seasonal effect needed: only if priceshad increased more than 20 percent would poverty have increased across a largerange of initial poverty levels and corresponding lines. Since both the CSA andERHS sources suggest a rural price inflation rate in this period of about 5-6 percent,the result on the seasonal decline in poverty is unlikely to be affected by theuncertainty about the actual inflation figures in this period. However, the changebetween 1994a and 1995 is more sensitive. At most levels of initial poverty, inflationrates below 10 percent would have implied declines in poverty. Since the CPIestimates inflation at only 3 percent, the use of these data would have resulted inestimates of substantial poverty declines, but such declines are contradicted by thedata collected in the survey sites. In contrast, over a very large range of povertylines, the prediction of 13 percent inflation as in the ERHS-survey would beconsistent with little change in poverty between 1994 and 1995.

7. Decomposing poverty changes

The FGT-poverty measures used in the analysis are additively decomposable, i.e.they can be written as a weighted average of poverty measures for subgroups, theweights being proportional to the population shares (Foster et al. (1984)). Formally,for m different subgroups, the poverty measures can be written as:

Pα=∑i=1m wi Pαi (5)

in which wi is the population share of subgroup i and Pαi is the poverty measure for

the subgroup. This property carries over to changes in poverty as well. Let s and t

21

be two periods in poverty measures are calculated and let (for simplicity) wi beconstant over time. Consequently, it follows from (5) that

Pαt - Pα

s =∑mi=1 wi (Pαti- Pαsi) (6)

It is then also possible to define θαi, the contribution of each group to the change inpoverty between t and s, as:

θαi = wi.(Pαti- Pαsi)/ (Pαt - Pαs) (7)

If θαi is larger (smaller) than wi, then the subgroup i has experienced proportionatelylarger (smaller) changes in poverty than the total population. As a first step in theanalysis of the dynamics of poverty, this is a useful statistic. Obviously, it is just astart and the interpretation suffers from all the problems static poverty profiles sufferfrom (Ravallion (1996)). The results must be regarded as descriptive statistics.

Applying this decomposition to the data from Ethiopia, we focus on a fewcharacteristics of the endowments of the households in the sample, which can beconsidered fixed in the short period under consideration. First, we look at somehuman capital variables in the broad sense: education (whether the head hascompleted primary school) and some labour supply characteristics of the head of thehousehold (the sex and the age of the head). Next, we consider some physicalassets: land owned in hectares and whether the household owns any oxen (or bulls).The former can be treated as exogenous to the household: land is not privatelyowned and is allocated by the peasant association to the household. Oxen are crucialin the main farming system for ploughing and cattle in general are an importantsource of wealth for accumulation. Of course, since markets exist, oxen ownershipmay well change over time, although the accumulation is generally slow. For thepoverty profile below, we use the ownership of oxen in 1994a31 . Finally, we look atsome infrastructure and location variables. As discussed before, there are somecritical differences in the experience of certain villages and village-level variablesmay well account for this. We grouped villages according to the distance to nearestall-weather road and to the distance to the nearest town.

We look at the contribution of different groups to changes over threeperiods: first, the change between 1989 and 1994-95 for the core panel villages,secondly, the change within 1994 allowing for some assessment of the sensitivity toseasonal variation and thirdly, the change between 1994 and 1995. For the first, weuse the food poverty measures and use food poverty in 1994-95 in the equivalentperiod of the data collection in 1989. We group the households in two (using themedian value for continuous variables32 ). We provide a t-test of the changes inpoverty for each subgroup and the contribution of each subgroup to the totalpoverty change. The relevant total changes for the full sample are given in tables 6and 8.

31 The analysis of the dynamics of oxen ownership in relation to poverty is beyond the scope of thispaper.32 For land, we considered both a grouping according to median land per village and according tomedian land per adult in the entire sample. If there are large differences in fertility, climate andfarming systems across villages in the sample, then the results may have been sensitive to thealternative groupings. However, the results were very similar, so we only report the results relativeto the median of land per adult in the entire sample.

22

Tables 9 a) to 9 g) give the results of the decompositions for the changesbetween 1989 and 1994/95 for the panel villages. Recall that poverty declined byabout 15.4 percentage points in this period (table 8). Human capital variables matterin accounting for the changes in this period. Although very few heads of householdare educated, they contributed proportionately more to the poverty decline.Similarly, households with younger heads experienced a larger decline in povertythan those with an older head of the household; the decline for the latter notsignificant even for the head count index. The sex of the household also matters:female headed households experienced no significant decline in this period. Oxenand land ownership is also important: those owning oxen and those with relativelylarge land holdings contributed proportionately more to the decline in poverty.Landholdings particularly affect the poverty gap and the severity of povertymeasure. The decline in poverty for those not owning oxen is not significantlydifferent from zero at 5 percent for all measures. Finally, distance to roads and totowns also matters a lot. At least with the respect to the head count index, thoseclose to all-weather roads contributed proportionately more to the decline inpoverty. Those households living more than 10 km outside towns experienced nosignificant change in poverty; consequently, the entire decline between 1989 and1994/95 can be accounted for by those in villages in the vicinity of urban areas.

Table 9 Decomposing changes in food poverty by sub-groups 1989-1994/95(n=351)

a) EducationHousehold head did not completeprimary school (97%)

Household head completed primary school (3%)

contrib contrib.Poverty89

poverty94

change t-test to change poverty89

poverty94

change t-test to change

P0 0,60 0,47 -0,14 -3,65** 87% 0,83 0,25 -0,58 -3,54** 13%P1 0,29 0,19 -0,10 -4,74** 90% 0,42 0,11 -0,30 -3,38** 10%P2 0,17 0,10 -0,07 -4,59** 93% 0,22 0,06 -0,15 -2,61** 7%

b) Age of the household headHead of the household is at least45 years(45%)

Head of the household is below 45years (55%)

contrib.poverty89

poverty94


poverty94

change t-test contrib. tochange

P0 0,59 0,50 -0,09 -1,84 33% 0,64 0,41 -0,23 -4,18** 67%P1 0,27 0,20 -0,08 -2,67** 39% 0,31 0,16 -0,15 -4,81** 61%P2 0,16 0,11 -0,06 -2,70** 42% 0,19 0,09 -0,10 -4,34** 58%

23

c) Sex of the head of the householdFemale headed household(17%)

Male headed household(83%)

contrib. contrib.Poverty89

poverty94


poverty94


P0 0,57 0,48 -0,09 -0,93 9% 0,62 0,45 -0,17 -4,12** 91%P1 0,29 0,20 -0,10 -1,85 15% 0,29 0,18 -0,11 -4,84** 85%P2 0,18 0,11 -0,07 -1,79 15% 0,17 0,10 -0,08 -4,55** 85%

d) oxen ownershipHousehold does not own oxen(33%)

Household owns at least one oxen(67%)

contrib. contrib.Poverty89

poverty94


poverty94


P0 0,65 0,55 -0,10 -1,62 22% 0,59 0,42 -0,18 -3,93** 78%P1 0,32 0,25 -0,08 -1,92 23% 0,28 0,15 -0,12 -5,13** 77%P2 0,20 0,14 -0,05 -1,82 24% 0,16 0,08 -0,08 -4,89** 76%

e) land ownershipLarge land holdings(50%)

Small land holdings (50%)

(above 0.45 ha) contrib. (below 0.45 ha) contrib.Poverty89

poverty94


poverty94


P0 0,54 0,38 -0,17 -3,15** 54% 0,68 0,54 -0,14 -2,76** 46%P1 0,27 0,13 -0,14 -5,26** 66% 0,32 0,24 -0,07 -2,38** 34%P2 0,16 0,06 -0,10 -5,36** 68% 0,19 0,14 -0,05 -2,04* 32%

f) distance to all-weather roadAt least 5 km from all-weather road(56%)

Less than 5 km from all-weather road (44%)

contrib. contrib.poverty89

poverty94


poverty94


P0 0,67 0,59 -0,08 -1,67 30% 0,54 0,29 -0,25 -4,54** 70%P1 0,34 0,25 -0,10 -3,35** 50% 0,23 0,11 -0,12 -4,39** 50%P2 0,21 0,13 -0,08 -3,61** 59% 0,13 0,06 -0,07 -3,53** 41%

24

g) distance to nearest townAt least 10 km from town(47%)

Less than 10 km fromtown(53%)

contrib. contrib.poverty89

poverty94


poverty 94 change t-test to change

P0 0,52 0,58 0,06 1,11 -19% 0,70 0,35 -0,34 -7,08** 119%P1 0,23 0,25 0,02 0,50 -7% 0,34 0,13 -0,22 -7,99** 107%P2 0,14 0,14 0,00 0,13 -2% 0,21 0,06 -0,14 -7,04** 102%**=significant at 1 percent*=significant at 5 percent

Turning to the larger sample, table 6 showed that poverty declined substantiallybetween the two rounds in 1994 (1994a and 1994b), but between 1994a and 1995,two rounds collected at roughly the same point in the seasonal cycle, povertyremained unchanged. The change within 1994 clearly reflects seasonal fluctuations.Table 10 shows which groups are more affected by these fluctuations. First, notethat households with older, female or uneducated heads have higher poverty levels inboth periods. However, the gap in poverty levels becomes smaller in 1994b. Forexample, those households whose heads have completed primary educationconstitute only one-tenth of the sample, but they did not experience any significantfluctuation, i.e. the entire decline in poverty is experienced by those householdswithout educated heads. Female-headed households while constituting about 22percent of the sample, contributed to about 40 percent of the change in the povertymeasures between 1994a and 1994b.

Higher asset ownership in terms of land and oxen, and distances to roads ortowns implies consistently lower poverty levels - but also accompanies largerfluctuations in consumption33 . Living closer to all-weather roads or towns is alsocorrelated with lower poverty in both periods, but linked to lower fluctuations. Inparticular, those living further away from towns contribute the lion’s share of thetotal poverty decline in this period. This may suggest that access to infrastructureand markets allows households to better smooth consumption. These are clearlyissues that need to be researched further.

33 This appears partly linked to the fact the fact that households with smaller land holdings areoften specialising more in permanent food crops such as enset, which provide a more stable returnover the season, which may help them to keep relatively smooth consumption.

25

Table 10 Decomposing changes in poverty by sub-groups 1994a-1994b (n=1411)

a) Education of the head of the householdHousehold head did not completeprimary

Household head completed primary school

school (91%) contrib. (9%) contrib.Poverty94a

poverty94b

change t-test to change poverty94a

poverty94b


P0 0,35 0,28 -0,08 -4,26** 99% 0,20 0,20 -0,01 -0,16 1%P1 0,14 0,08 -0,05 -6,62** 101% 0,05 0,06 0,00 0,26 -1%P2 0,07 0,04 -0,04 -6,68** 101% 0,02 0,03 0,00 0,39 -1%

b) Age of the household headHead of the household is at least 45years(52%)

Head of the household is below 45years(48%)

contrib. contrib.poverty94a

poverty94b

change t-test tochange

poverty94a

poverty94b


P0 0,39 0,31 -0,07 -2,91** 52% 0,29 0,22 -0,07 -3,00** 48%P1 0,15 0,09 -0,06 -5,11** 60% 0,11 0,07 -0,04 -3,90** 40%P2 0,08 0,04 -0,04 -5,20** 61% 0,06 0,03 -0,03 -3,90** 39%

c) Sex of the head of the householdFemale headed household(22%)



poverty94b

change t-test tochange

poverty94a

poverty94b


P0 0,40 0,28 -0,12 -3,17** 37% 0,33 0,27 -0,06 -2,99** 63%P1 0,17 0,08 -0,09 -4,93** 40% 0,12 0,08 -0,04 -4,52** 60%P2 0,10 0,04 -0,06 -4,84** 40% 0,06 0,04 -0,02 -4,63** 60%

d) Oxen ownershipHousehold does not own oxen (48%) Household owns at least one oxen

(52%)contrib.

poverty94a

poverty94b

change t-test contrib.to change

Poverty94a

poverty94b


P0 0,38 0,34 -0,04 -1,59 28% 0,30 0,20 -0,10 -4,37** 72%P1 0,15 0,11 -0,03 -2,87** 34% 0,12 0,05 -0,06 -6,60** 66%P2 0,08 0,05 -0,02 -3,08** 38% 0,06 0,02 -0,04 -6,71** 62%

e) Land ownershipLarge land holdings (50%) Small land holdings (50%)(above 0.23 ha) contrib. (below 0.23 ha) contrib.poverty94a

poverty94b


poverty94b


P0 0,28 0,16 -0,11 -5,22** 80% 0,40 0,38 -0,03 -1,10 20%P1 0,10 0,04 -0,06 -6,46** 58% 0,16 0,12 -0,04 -3,34** 42%P2 0,05 0,02 -0,03 -6,02** 50% 0,09 0,06 -0,03 -3,90** 50%

26

f) Distance to the nearest all-weather roadAt least 5 km from all-weather road(44%)



poverty94b


poverty94b


P0 0,47 0,36 -0,11 -4,09** 69% 0,24 0,20 -0,04 -1,87 31%P1 0,20 0,11 -0,09 -6,61** 77% 0,08 0,06 -0,02 -2,33** 23%P2 0,11 0,05 -0,06 -6,78** 80% 0,04 0,03 -0,01 -2,11* 20%

g) Distance to the nearest townAt least 10 km from town(45%)

Less than 10 km from town (55%)


poverty94b


poverty94b


P0 0,38 0,24 -0,14 -5,44** 88% 0,31 0,29 -0,02 -0,67 12%P1 0,15 0,08 -0,07 -6,17** 68% 0,11 0,08 -0,03 -2,88** 32%P2 0,08 0,04 -0,05 -5,76** 65% 0,06 0,04 -0,02 -3,34** 35%**=significant at 1 percent*=significant at 5 percent

Table 11 gives the decompositions comparing the first round of 1994 with the 1995data. Since overall the change is insignificant, the contribution to the percentagechange in the table are not relevant. The table shows that for none of the groupsdefined by education, age of the head, the sex of the head or oxen ownership is therea significant change between 1994a and 1995. Those with relatively more land saw afurther fall in poverty in this period. In terms of the higher order poverty measures,those living further from all-weather roads also improved their situation. The lattergroup is however still much poorer than those living near to roads and had not beenable to benefit as much in the period 1989-1994, so if anything this effect suggestssome limited catching up. The changes between 1994a and 1995 remain small, sothe conclusion that relatively little changed in this period stands.

Table 11 Decomposing changes in poverty by sub-groups 1994a-1995 (n=1411)

a) Education of the head of the householdHousehold head did not completeprimary school (91%)

Household head completed primary school (9%)


poverty95


poverty95


P0 0,35 0,37 0,01 0,74 100% 0,20 0,20 0,00 0,00 0%P1 0,14 0,14 0,00 0,24 84% 0,05 0,06 0,00 0,23 16%P2 0,07 0,07 0,00 -0,23 153% 0,02 0,03 0,00 0,48 -53%

27

b) Age of the household headHead of the household is at least45 years(52%)

Head of the household is below 45years(48%)


poverty95


poverty95


P0 0,39 0,42 0,03 1,33 139% 0,29 0,28 -0,01 -0,42 -39%P1 0,15 0,16 0,01 1,15 317% 0,11 0,10 -0,01 -0,95 -217%P2 0,08 0,08 0,01 0,73 -385% 0,06 0,05 -0,01 -1,18 485%

c) Sex of the household headFemale headed household(22%)



poverty95


poverty95


P0 0,40 0,37 -0,03 -0,66 -44% 0,33 0,35 0,02 1,17 144%P1 0,17 0,14 -0,03 -1,49 -274% 0,12 0,13 0,01 1,21 374%P2 0,10 0,07 -0,02 -1,63 574% 0,06 0,07 0,00 0,83 -474%

d) Oxen ownershipHousehold does not own oxen(48%)

Household owns at least one oxen(52%)


poverty95


poverty95


P0 0,38 0,42 0,04 1,44 153% 0,30 0,29 -0,01 -0,52 -53%P1 0,15 0,17 0,02 1,73 612% 0,12 0,10 -0,02 -1,66 -512%P2 0,08 0,09 0,01 1,30 -529% 0,06 0,05 -0,01 -1,93 629%

e)Land ownershipLarge land holdings (50%) Small land holdings (50%)(above 0.23 ha) contrib. (below 0.23 ha) contrib.poverty94a

poverty95


poverty95


P0 0,28 0,23 -0,05 -2,07* -189% 0,40 0,48 0,07 2,81** 289%P1 0,10 0,07 -0,03 -3,13** -635% 0,16 0,20 0,03 2,56** 735%P2 0,05 0,03 -0,02 -3,35** 1142% 0,09 0,11 0,02 1,86 -1042%

f) Distance to the nearest all-weather roadAt least 5 km from all-weather road(44%)



poverty95


poverty95


P0 0,47 0,44 -0,03 -0,92 -89% 0,24 0,28 0,04 1,93 189%P1 0,20 0,16 -0,03 -2,15* -562% 0,08 0,11 0,03 2,79** 662%P2 0,11 0,08 -0,02 -2,57* 1279% 0,04 0,06 0,02 2,70** -1179%

28

g) Distance to the nearest townAt least 10 km from town(45%)

Less than 10 km from town (55%)


poverty95


poverty95


P0 0,38 0,36 -0,02 -0,70 -67% 0,31 0,34 0,04 1,63 167%P1 0,15 0,14 -0,01 -0,77 -197% 0,11 0,12 0,01 1,18 297%P2 0,08 0,07 -0,01 -1,00 488% 0,06 0,06 0,01 0,85 -388%**=significant at 1 percent*=significant at 5 percent

7. Conclusions

This paper investigates the problems of comparing poverty over time in a panelhousehold survey collected between 1989 and 1995 in rural Ethiopia. We used FGT-measures of poverty and implemented significance tests for changes in poverty. Wefound that poverty declined substantially between 1989 and 1995. Poverty remainedlargely unchanged between 1994 and 1995. We found substantial differences inpoverty levels between the two rounds of data collection in 1994, suggestingsubstantial seasonal fluctuations.

These results were found to be robust, firstly, to changes in the definitions ofconsumption used and to small changes in questionnaire design. These factorsaffected the actual magnitudes involved, but not the overall thrust of the findings.Secondly, using stochastic dominance tests, the changes were found to persist acrossdifferent poverty lines. Thirdly, the actual poverty measures were sensitive toseasonal factors: as the results for 1994 showed, the exact timing of the datacollection matters for the magnitudes of the poverty measures. This point is rarelyconsidered when comparing poverty over time in developing countries, especiallysince data collection usually takes many months to be completed which implies thatconsumption changes are not readily comparable. Still, this problem, while importantin general, did not affect the main conclusions about poverty changes between 1989and 1995 in our sample. Fourthly, there appears to be substantial inconsistencybetween different sources of price data in assessing the appropriate rural cost-of-living deflators for consumption between the different periods. Testing therobustness of the results to the uncertainty surrounding the cost-of-living deflators isnot self-evident, since it is not possible using the standard dominance tests. Wesuggest and implement a simple graphical technique to do so. The results for theobserved changes between 1989 and 1995 are clearly robust to any reasonableinflation estimate. The same applies to the changes within 1994. For the changesbetween 1994 and 1995, using the official inflation figures would yield a decline inpoverty, but using the data collected in the specific rural areas would support theconclusion of unchanged poverty.

In a final section, we provide a simple decomposition of the findings acrossdifferent groups. We found that those with relatively better human capital or laboursupply characteristics (better education, male headed households and relativelyyoung heads of households) experienced levels of poverty in each period. They alsohad larger poverty declines between 1989 and 1995 and lower seasonal fluctuationsin poverty. Households with better physical capital endowments, in terms of land

29

and oxen, had lower poverty levels and saw larger poverty declines. They seem toface larger poverty fluctuations across seasons. Finally, those with good access toroad infrastructure and those close to towns also had lower poverty levelsthroughout. They experienced a larger poverty decline between 1989 and 1995 andexperienced lower within-year fluctuations.

These decomposition results are only the first step in the analysis of thedynamics of poverty. Nevertheless, the results appear to suggest that not only didphysical, human and infrastructural capital matter in explaining levels of poverty, butthat poverty declines during a period of reforms and the return to peace appear to beinfluenced by the initial levels of these sources of capital, so that better-endowedhouseholds were better placed to benefit much more from the changedcircumstances. Similarly, access to infrastructure and proximity to towns, as well asbetter human capital circumstances, implies lower fluctuations in seasonal povertylevels, presumably linked to more opportunities for alternative income generationand smaller food price fluctuations. Clearly, these issues need to be investigatedfurther. Note also that we while we did use panel data, we did not exploit its fullpotential34 . In fact, the entire analysis, including the robustness tests

changes in poverty in rural ethiopia 1989-1995 ... · the last problem of using appropriate price...

Documents